Eye on AI Weekly Research Watch

Епизоди

• 
  Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations 19.08.2026 2мин

  Authors: Toby D. Pilditch Authors: Toby D. Pilditch Paper: https://arxiv.org/abs/2608.14425v1
• 
  The Dynamics of Intelligence Explosions 19.08.2026 2мин

  As AI increasingly accelerates its own research and development, could this create a runaway feedback loop of self-improvement? This theoretical paper mathematically analyzes the conditions required for such an "intelligence explosion," showing that truly explosive, asymptotic growth is harder to achieve than prior economics-based models suggested, while identifying an underappreciated middle category of faster-than-exponential (but bounded) growth. The paper highlights "generation time" --- how long one loop of the feedback cycle takes --- as the crucial hidden variable. This work matters for AI safety, forecasting, and policy discussions about transformative AI timelines. Authors: Toby Ord Paper: https://arxiv.org/abs/2608.14426v1
• 
  PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments 19.08.2026 2мин

  Self-improving AI agents are usually tested under fixed conditions, but real-world deployment demands adapting when the environment itself changes. PACE-Bench introduces 144 source-to-target adaptation challenges across six physics domains, forcing agents to iteratively rewrite working code when the underlying physics is mutated. Testing ten methods reveals that grounded, feedback-driven revision beats memory-based or unguided search, and that even knowing the exact physical change doesn't guarantee success --- redesigning the approach matters more than fine-tuning parameters. This benchmark is valuable for evaluating robust, adaptable AI agents for robotics and simulation. Authors: Yuhao Zhan, Bingxiang He, Zecong Tang, Chaojun Xiao Paper: https://arxiv.org/abs/2608.14441v1
• 
  Designing Compact Neural Architectures via Neuron Gating and Mixed Activation 19.08.2026 2мин

  Neural Architecture Search is powerful but expensive due to discrete, combinatorial design choices. This paper proposes continuous relaxations of neuron-level and activation-level decisions, enabling fully differentiable optimization across MLPs, CNNs, RNNs, and Transformers. Three resulting methods (NAS-NG, NAS-MA, NAS-NGMA) find highly compact architectures --- including a CNN with just 0.26M parameters hitting 99.63% MNIST accuracy --- while outperforming standard DARTS on CIFAR-10. This offers a scalable, general-purpose toolkit for automatically designing efficient models, valuable for deploying AI on resource-constrained devices like mobile phones or edge hardware. Authors: Abhishek Shukla, Ankur Sinha, Faiz Hamid Paper: https://arxiv.org/abs/2608.14443v1
• 
  Wyvern: An Agentic Framework for Generating Grounded Multimodal Reports 19.08.2026 2мин

  As AI-generated content proliferates faster than anyone can verify it, ungrounded generative reports risk spreading unsupported claims. Wyvern addresses this with a multi-agent framework that produces technical reports combining text, images, and tables, all tied to supporting references, with a dedicated claims-revision stage to strengthen grounding. Human evaluators rated Wyvern's figures as more informative than baselines in 87% of cases, and its reports as more useful in the majority of comparisons. This has clear applications in automated research synthesis, technical documentation, journalism, and any domain needing trustworthy AI-generated multimodal reporting. Authors: Beatrice Alessandra Motetti, Emilien Guandalino, Daniele Jahier Pagliari, Alessio Burrello, Lorenz K. Müller, Konstantin Berestizshevsky, Lukas Cavigelli Paper: https://arxiv.org/abs/2608.14446v1
• 
  SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning 19.08.2026 2мин

  Spreadsheets pack in complex, spatial, and cross-referential structure that gets destroyed when LLMs flatten them into plain text, crippling automated analysis. SheetCompass instead builds explicit relation graphs capturing structural links within and across worksheets, paired with a memory system that keeps relevant context available to reasoning agents. This lets AI agents navigate spreadsheets more like human analysts do --- spatially and relationally. Applications include automated financial modeling, business intelligence tools, spreadsheet auditing, and any enterprise workflow where AI agents need to reliably interpret and manipulate real-world Excel files. Authors: Panjing He, Mingyue Cheng, Yucong Luo, Li Li, Xiaohan Zhang Paper: https://arxiv.org/abs/2608.14452v1
• 
  Shift Aware Transfer Learning with Adaptive Dual-Encoder Fusion for PM Forecasting in Data-Limited Environments 19.08.2026 2мин

  Forecasting air pollution (PM2.5) is hard where local sensor data is scarce, and naive transfer learning from data-rich regions can actually hurt performance due to domain mismatch. This study's dual-encoder framework pretrains on U.S. monitoring data, then adaptively fuses it with limited Taiwan-specific data, letting the source model adjust rather than freeze. The adapted model beat baselines significantly, with SHAP analysis confirming recent pollution levels and weather as key drivers. This has direct applications for environmental monitoring agencies in developing regions with sparse air quality infrastructure. Authors: Shahab Band, Hamed Mohammadi Paper: https://arxiv.org/abs/2608.14456v1
• 
  LP-NAS: Linear Programming-based Neural Architecture Search 19.08.2026 2мин

  Automating neural network design (NAS) is powerful but computationally costly, and differentiable NAS methods like DARTS often converge slowly. LP-NAS reframes the architecture search as a linear programming problem, using gradient and Hessian information to compute better-informed update directions that improve generalization while preserving optimized parameters. Its two variants integrate into the popular DARTS framework and show faster, stronger convergence on CIFAR-10/100, with transferability to ImageNet. This offers machine learning practitioners a more efficient, mathematically grounded alternative for automating model design across computer vision and beyond. Authors: Abhishek Shukla, Ankur Sinha, Faiz Hamid Paper: https://arxiv.org/abs/2608.14472v1
• 
  Ensuring Safe Physical AI in Urban Mobility via Hazard-Informed Synthesized Envelopes 19.08.2026 2мин

  As robots increasingly share city streets and sidewalks with people, ensuring safety can't just be handled by isolated software modules --- it needs to span the whole system. This paper proposes hazard-informed safety envelopes that connect systematic hazard analysis to real-time runtime enforcement, unifying symbolic, spatial, and dynamic representations of the environment. The framework is designed to plug directly into physical AI runtime harnesses. Applications include delivery robots, autonomous urban vehicles, and any human-robot interaction system operating in unpredictable public spaces where safety guarantees are critical. Authors: Alexei Odinokov, Rostislav Yavorskiy Paper: https://arxiv.org/abs/2608.14481v1
• 
  Twin: Playing an Unknown Game with a Test-Time Digital Twin 19.08.2026 2мин

  Instead of hand-engineering a world model for each new task, this paper has a coding agent write executable world models on the fly, purely from observing an unfamiliar game. A "twin" simulation validates each hypothesis by requiring it to reproduce every past observation before acting, using mismatches to iteratively repair the model. Tested on ARC-AGI-3 games, Twin dramatically outperforms baseline agents and even exceeds human efficiency on most levels. This approach has broad implications for general-purpose game-playing AI, robotics, and any domain requiring rapid, autonomous understanding of unknown environments. Authors: Alexy Skoutnev, Kirill Acharya, Gaston Longhitano, Madeleine Udell, Kevin Ellis, Iddo Drori Paper: https://arxiv.org/abs/2608.14490v1
• 
  Optimal Scheduling of Road Maintenance Jobs Considering Impact on Traffic Flows 19.08.2026 2мин

  Planning road maintenance at scale requires repeatedly simulating how traffic redistributes when capacity is reduced --- a process that's normally too computationally expensive to embed directly into scheduling optimization. This paper builds data-driven surrogate models that predict equilibrium traffic flows directly from origin-destination demand, bypassing expensive equilibrium solvers, and validates the approach using real traffic data from Newark, NJ. This offers city planners and transportation departments a scalable tool for smarter maintenance scheduling that minimizes disruption, with potential extension to broader infrastructure and network planning problems. Authors: Charitha Nandepu, Lohitha Kalepu, Gabriele Ciavarella, SangWoo Park Paper: https://arxiv.org/abs/2608.14491v1
• 
  Generating Benchmark Health Data Using a Tabular Diffusion Transformer 19.08.2026 2мин

  Healthcare researchers often need synthetic data spanning multiple related but heterogeneous tables, yet most generative models only handle single tables. This paper's two-stage framework first standardizes diverse tables into common statistical summaries capturing distributions and correlations, then uses a diffusion transformer to generate new synthetic statistical tables, which are reconstructed back into realistic raw data. This enables privacy-preserving generation of unlimited realistic multi-table datasets. Applications include healthcare data sharing without exposing patient records, benchmark creation for medical AI research, and synthetic data generation for any domain with complex relational tables. Authors: Hao Yan, Lisa Pilgram, Dan Liu, Linglong Kong, Fida Dankar, Khaled El Emam Paper: https://arxiv.org/abs/2608.14496v1
• 
  Universal Thermodynamic Interatomic Potentials for Crystalline Materials 19.08.2026 2мин

  Discovering new materials computationally usually relies on ground-state energy calculations, ignoring temperature and pressure effects that determine real-world phase stability. This paper introduces TIP, which upgrades an interatomic potential model into a full thermodynamically consistent free energy model using automatic differentiation, letting a single calculation reveal phase transitions and equations of state. Built on the universal UMA potential and fine-tunable to alloys, TIP makes finite-temperature behavior as computationally accessible as static energy. This has major implications for high-throughput materials discovery, alloy design, and computational chemistry pipelines. Authors: Juno Nam, Bowen Deng, Xiaochen Du, Luis Barroso-Luque, Benjamin Kurt Miller, Rafael Gómez-Bombarelli Paper: https://arxiv.org/abs/2608.14502v1
• 
  RecipeNet: A Hierarchical Transformer for Recipe Data 19.08.2026 2мин

  Many real-world processes --- chemical synthesis, drug formulation, manufacturing --- are naturally represented as ordered, structured steps rather than flat tables, yet most machine learning models flatten this structure and lose crucial dependencies. RecipeNet addresses this with a hierarchical Transformer that separately models relationships within each step and dependencies across the full sequence. By outperforming standard tabular learning methods, it offers a general-purpose architecture for procedural or "recipe-like" data. Potential applications include materials science R&D, pharmaceutical formulation optimization, and industrial process modeling wherever step-ordered structured data needs to be learned. Authors: Pin-Yen Huang, Sachin Chhabra, Prasanth Sai Gouripeddi, Abhinav Kumar, Baoxin Li Paper: https://arxiv.org/abs/2608.14505v1
• 
  Split the Labor: Separating Evidence Interpretation from Decision Aggregation 19.08.2026 2мин

  When LLMs are asked to synthesize conclusions from many sources, cramming everything into one prompt conflates two different jobs: understanding each piece of evidence and combining conclusions fairly. This paper identifies a subtle failure mode, "count-scale drift," where naive vote-counting shifts as more sources are added, and proposes separating interpretation from aggregation using a structured evidence format and calibrated probability pooling. Validated on a real medical/survival dataset, this approach could improve any multi-source reasoning system, from diagnostic AI panels to fraud detection to automated research synthesis tools. Authors: Zhelun Wu Paper: https://arxiv.org/abs/2608.14509v1
• 
  Learning-to-Transition for Large-scale and High-Order MIMO Detection 19.08.2026 2мин

  Modern wireless systems need to decode signals sent simultaneously from many antennas, a computationally brutal search problem especially at high signal orders. This paper reframes MIMO detection as a learned sequential decision process, using a Transformer to guide symbol-by-symbol search while producing reliable confidence scores for downstream error correction. A clever "hard-to-soft" training trick transfers a well-trained detector into a full iterative decoding pipeline. Applications span 5G/6G wireless infrastructure, massive MIMO systems, and any communications hardware needing fast, accurate, learnable alternatives to traditional signal detection algorithms. Authors: Yubo Zhang, Yiyao Liu, Xiaodong Wang Paper: https://arxiv.org/abs/2608.14511v1
• 
  Participatory Moral AI Is Not Neutral: The Invisible Hand of Developers 19.08.2026 2мин

  Democratic AI alignment approaches poll people on moral dilemmas to train policies --- but this study shows the "neutral" aggregation process is shaped by hidden developer choices. Across studies involving kidney allocation, AI worker simulation, and deepfakes of the deceased, the researchers show that which features get voted on, who votes, and how questions are worded all systematically shift outcomes, sometimes along political lines. This has major implications for AI governance and alignment research, arguing that voting-based systems need transparency audits at every design stage, not just at the final vote tally. Authors: Taenyun Kim, Edyta Bogucka, Daniele Quercia Paper: https://arxiv.org/abs/2608.14522v1
• 
  Handover of In-Context Learning State Across Session Boundaries 19.08.2026 1мин

  When an LLM conversation exceeds context limits, restarts, or gets handed to another agent, someone must decide what information survives the transition. This paper formalizes that handover problem mathematically, distinguishing exact recovery from statistically sufficient preservation, and proposes a three-part memory record for decisions, summarized evidence, and irreplaceable raw observations. With theoretical bounds from Gaussian regression and nonparametric settings, it quantifies how much memory a task-continuation actually requires. This is directly applicable to multi-agent LLM systems, long-running assistants, and any application needing principled memory compression across session boundaries. Authors: Masahiro Kato, Taka Kato Paper: https://arxiv.org/abs/2608.14528v1
• 
  Marionette: Predicting World States, Rendering Geometry, Painting Appearance 19.08.2026 2мин

  Game world models that generate video directly often lose track of physical consistency over long sequences --- characters clip through walls or drift apart. Marionette fixes this by separating concerns: a neural model predicts explicit 3D skeletal states, a zero-parameter renderer handles geometry exactly, and video diffusion paints the final appearance. This structure lets developers impose simple rules (like collision limits) directly on the state to fix long-horizon errors, without retraining the visual model. Applications include interactive game engines, character animation systems, and any simulation needing controllable, physically coherent long-horizon generation. Authors: Zian Meng, Zhen Li, Chuanhao Li, Qiang Li, Kaipeng Zhang Paper: https://arxiv.org/abs/2608.14530v1
• 
  Decoding the Past: An Uncertainty-Aware Deep Learning Framework for Sex Attribution in Prehistoric Hand Stencils 19.08.2026 2мин

  Archaeologists have long puzzled over who made Upper Paleolithic hand stencils, since no ground truth exists and prehistoric hands differ from modern ones. This paper tackles that uncertainty head-on with a deep learning pipeline that generates multiple silhouette interpretations per stencil, then runs them through two neural network ensembles trained on thousands of contemporary hand images. By combining this with manifold mapping and explainable AI, the system doesn't just guess sex --- it flags how confident that guess is. Applications include rock art archaeology, anthropology, and any forensic-style problem where degraded imagery demands uncertainty-aware classification rather than false certainty. Authors: Karel Becerra, Boris Mederos, Dean Snow, Ramón A. Mollineda Paper: https://arxiv.org/abs/2608.14539v1